PrismML Compresses Qwen3.6 to 4GB, Fits 27B Model into iPhone

According to Dynamic Beating monitoring, PrismML, an AI lab co-founded by Caltech mathematician Babak Hassibi, has successfully run the Alibaba open-source model Qwen3.6-27B compressed on an iPhone 17 Pro.

Qwen3.6-27B is a 27-billion-parameter dense model where all parameters are involved in inference, unlike MoE models that only activate certain parameters.

PrismML stated that the model size has been compressed from around 54GB to less than 4GB with no significant loss in performance. It can be used for complex dialogues, inference, agents, and code-related tasks.

Apple has been in touch with PrismML to discuss the implementation of this technology. Apple has been looking to incorporate more AI locally on devices, but advanced Siri still relies on cloud-based models.